scRNAseq clustering tools

Åsa Björklund
[email protected]
 Cell identity

Wagner et al. Nat. Biotech 2016

 What is a cell type?

• A cell that performs a specific function?

• A cell that performs a specific function at a specific
location/tissue?
• Not clear where to draw the line between cell types
and subpopulations within a cell type.
• Also important to distinguish between cell type and
cell state.
– A cell state may be infected/non infected
– Metabolically active/inactive
– Cell cycle stages
– Apoptotic
 Outline

• Basic clustering theory

• Graph theory introduction
• scRNAseq clustering with graphs
• Examples of different tools
How can we identify populations?
 Considerations for clustering

• Hypotheses:
– What is a cell type? What cell types are in my tissue?
– What is the number of clusters k?
• Choices:
– Gene set selection
– Similarity measure / Space to calculate similarity
– Algorithm and hyper parameters of that algorithm.
• Different choice leads to different results. Validate,
interpret and repeat steps.
 What is clustering?

• “The process of organizing objects into groups whose

members are similar in some way”
• Typical methods are:
– Hierarchical clustering
– K-means clustering
– Density based clustering
– Graph based clustering
 The main idea

• Structure when:
1) Samples within cluster
resemble each other (within
variance, σW(i))
2) Clusters deviate from each
other
(between variance, σB)
Group samples such that:
 Hierarchical clustering

• Builds on distances between data points

• Agglomerative – starts with all data points as
individual clusters and joins the most similar ones in
a bottom-up approach
• Divisive – starts with all data points in one large
cluster and splits it into 2 at each step. A top-down
approach
• Final product is a dendrogram representing the
decisions at each merge/division of clusters
Hierarchical clustering
 Hierarchical clustering

Clusters are obtained by cutting the tree at a desired level

 Hierarchical clustering

Clusters are obtained by cutting the tree at a desired level

 Linkage criteria
• Calculation of similarities between 2 clusters (or a
cluster and a data point)

http://www.slideshare.net/uzairjavedsiddiqui/malhotra20
• Ward (minimum variance method). Similarity of two clusters is
based on the increase in squared error when two clusters are
merged.
http://www.slideshare.net/uzairjavedsiddiqui/malhotra20
 K-means clustering
1. Starts with random selection of cluster centers
(centroids)
2. Then assigns each data points to the nearest cluster
3. Recalculates the centroids for the new cluster
definitions
4. Repeats steps 2-3 until no more changes occur.
Can use same distance measures as in hclust.

https://en.wikipedia.org/wiki/K-means_clustering
 Network/graph clustering
Node/Vertice
Community

Edge –
(weighted
& directed)

Hubs

Connectivity
- # of edges

(http://www.lyonwj.com/2016/06/26/
graph-of-thrones-neo4j-social-network-analysis/)
 Types of graphs

• The k-Nearest Neighbor (kNN) graph is a graph in

which two vertices p and q are connected by an
edge, if the distance between p and q is among the
k-th smallest distances from p to other objects from
P.
• The Shared Nearest Neighbor (SNN) graph has
weights that defines proximity, or similarity between
two edges in terms of the number of neighbors (i.e.,
directly connected vertices) they have in common.
 SNN graph

(Ertöz et al. Semantic scholar, 2002)

 SNN graph

• A common measure of shared neighbors is the

Jaccard index:
– Shared neighbors / Total neighbors for both.
• Other measures includes rank, number of shared
neighbors, overlap coefficient
• Common to do pruning – remove all edges
between nodes with e.g. Jaccard similarity <
cutoff.
Graphs, adjacency and weight matrices
 Community detection
Communities, or clusters, are usually groups of vertices
having higher probability of being connected to each other
than to members of other groups.
 Community detection

• Main objective is to find a group (community) of

vertices with more edges inside the group than
edges linking vertices of the group with the rest of
the graph.
 Graph cuts

• Graph cut partitions a graph

into subgraphs
• Cut cost is the sum of weights
of the edges.
• Clustering by graph cuts: find
the smallest cut that bi-
partitions the graph
• The smallest cut is not always
the best cut – may give many
small disjoint cluster
 Normalized cut
• Normalized cut computes the cut cost as a fraction of
the total edge connections to all the nodes in the
graph.
 Normalized cut

• Searching for the best normalized cut is NP-hard

• We need a heuristic method to solve the problem:
– Spectral clustering
– Markov clustering
– Louvain
– Leiden
– ...
 For single cell data

• Can start with distances based on correlation,

euklidean distances in PCA space etc. Same as for
hclust/k-means.
• Buld a KNN graph with cells as vertices.
– Find k nearest neighbors to each cell.
– The size of k will strongly influence the network structure.
• Can create weighted network based on shared
neighbors (SNN).
• Find clusters with community detection method.
• Graphs can also be used for trajectory analysis etc.
 How to work with networks

• Igraph package – implemented for both R, python and Ruby

• Has most commonly used layout optimization methods and
community detection methods implemented.

• Simple R example at:

https://jef.works/blog/2017/09/13/graph-based-community-
detection-for-clustering-analysis/
• Tutorial to igraph at:
http://kateto.net/networks-r-igraph
• Example how to build your own graph with scRNAseq data:
https://github.com/NBISweden/workshop-scRNAseq/blob/master/oldlabs/igraph.md
 Distance between cells

• All clustering methods need to define distances

between cells. Things to consider are:
– What gene set should be included?
• Commonly used: Highly variable genes
– What space to calculate distance in?
• Commonly done in PCA space
• Can also be full space, tSNE, UMAP etc.
– How many dimensions to include?
– What distance measure?
 Different distance measures

• Most commonly used in scRNA-seq:

– Euclidean distance
– Inverted pairwise correlations (1-correlation)
• Other common methods are:
– Manhattan distance
– Mahalanobis distance
– Maximum distance
 Selection of principal components

• To overcome the extensive

technical noise in scRNA-seq
data, it is common to cluster
cells based on their PCA
scores
• Each PC represents a
‘metagene’ that (linearly)
combines information across
a correlated gene set
• Depending on the
heterogeneity of your data
more/less PCs should be
selected.
 Bootstrapping

• How confident can you be that

the clusters you see are real?
• You can always take a random
set of cells from the same cell
type and manage to split them
into clusters.
• Most scRNAseq packages do
not include any bootstrapping.
Scran has function
bootstrapCluster.

(Rosvall et al. Plos One 2010 )

 Seurat clustering

• FindNeighbors:
– First construct a KNN (k-nearest neighbor) graph – default is based on
the euclidean distance in PCA space
– Then SNN graph the edge weights between any two cells based on the
shared overlap in their local neighborhoods (Jaccard distance) and
pruning of distant edges.
• Important parameters:
– reduction: default is “pca”
– dims: number of PCs
– k.param: Number of neighbors in KNN graph
– prune.snn: Cutoff for pruning

(http://satijalab.org/seurat/)
 Seurat clustering

• FindClusters: To cluster the cells, modularity optimization:

– Louvain
– Louvain with multilevel refinement
– Leiden
– SLM
• Important parameters:
– resolution: default 0.8, larger values give more communities, smaller
gives less.
– Algorithm

(http://satijalab.org/seurat/)
 Scran clustering

• buildKNNGraph: Constructs the KNN graph

• buildSNNGraph: Constructs KNN and then SNN
graph. OBS! Adds weighted edges to cells that shares
neighbors.
– Allows for different similarity measures, default is “rank”
but also includes “jaccard” or “number”
• Important parameters:
– use.dimred: dimensionality reduction to use
– k: number of neighbors
– type: weighting method
 Scran clustering

• Community detection is done with the igraph

package.
– cluster_louvain
– cluster_leiden
– cluster_infomap
– Many more…
• No resolution parameters, can mainly tweak cluster
number with the k parameter.
 Scanpy clustering

• sc.pp.neighbors – creates KNN graph

– Has many different options for distance calculation, default
is euclidean.
– No SNN graph construction
• Clustering:
– sc.tl.leiden
– sc.tl.louvain
– Can specify resolution like in Seurat.
Pagoda – Pathway And Geneset OverDispersion Analysis
Implemented in the SCDE package

(Fan et al. Nature Methods 2016)

Pagoda – Pathway And Geneset OverDispersion Analysis

• Helps with biological interpretation of data

• Important to have good and relevant gene sets
• High memory consumption when running Pagoda
• Also has methods for removing batch effect, detected genes, cell cycle etc

(Fan et al. Nature Methods 2016)

 Pagoda2

• Similar error modelling

• Now include KNN graph clustering
• largeViz for dimensionality reduction
• Can visualize gene sets.
• https://github.com/hms-dbmi/pagoda2
 HDBSCAN

• Hierarchical DBSCAN – density based clustering on

tSNE / Umap
Loupe – Cell Browser, from 10X Genomics
 Which clustering method is best?

• Depends on the input data

• Consistency between several methods gives
confidence that the clustering is robust
• The clustering method that is most consistent – best
bootstrap values is not always best
• In a simple case where you have clearly distinct
celltypes, simple hierarchical clustering based on
euclidean or correlation distances will work fine.
Comparison of clustering methods
 How many clusters do you really have?

• It is hard to know when to stop clustering – you can

always split the cells more times.
• Can use:
– Do you get any/many significant DE genes from the next
split?
– Some tools have automated predictions for number of
clusters – may not always be biologically relevant
• Always check back to QC-data – is what your splitting
mainly related to batches, qc-measures (especially
detected genes)
 Clustree – R package

https://cran.r-project.org/web/packages/clustree/vignettes/clustree.html
 Subclustering

• Most of the variation in a heterogeneous data set

will be between broad celltypes.
• By selecting one celltype and rerunning HVG-
selection and PCA – most of the variation will be
differences between subtypes.
Check QC data
Check QC data
Check QC data
Check batches / conditions
 Considerations for clustering

• Hypotheses:
– What is a cell type? What cell types are in my tissue?
– What is the number of clusters k?
• Choices:
– Gene set selection
– Similarity measure / Space to calculate similarity
– Algorithm and hyper parameters of that algorithm.
• Different choice leads to different results. Validate,
interpret and repeat steps.
 Conclusions

• Clearly distinct celltypes will give similar results

regardless of method
• Subclustering within celltypes may require careful
selection of variable genes, dim reduction etc.
• Consistent results from different methods and
agreement with tSNE/UMAP layout is always best!
• Use your biological knowledge to evaluate the results
– but try to be unbiased!